Skip to main content
A conversation is one thread of work with the agent. It keeps its own history, runs in its own cloud session, and can belong to a project.

​
The composer

The composer is the box you type in: “Ask anything, or describe a task…”. Around it are the controls that shape a message.
Prefer Reference over Attach for anything already in a vault. A reference always points to the latest version, costs nothing to upload, and lets the agent edit the real document rather than a copy.

​
Context from your screen

With Include what you’re looking at on, “summarize this” and “draft a reply to this” work as you would expect. Up to 12 open tabs are included. If you have more, the composer tells you how many were left out so you can close some.

​
Modes

Type / and choose a mode.
The default. The agent reads, writes and acts to complete the task, within your permission level.

​
Skills in a message

Type / followed by a skill’s name, for example /draft-email or /property-due-diligence, to tell the agent to follow that playbook for this message. The agent also picks skills by itself when a task matches one. See Skills.

​
Memory

The agent remembers useful context about you between conversations, so that you do not start from nothing each time. Memory belongs to you, within one organization. It is not shared with colleagues, and if you belong to more than one organization, what the agent remembers in one never appears in another.
In conversations that other people can read, namely shared conversations and matter threads, memory is confined to that conversation. The agent neither draws on your private conversations nor updates your personal profile there, so nothing private surfaces in front of colleagues.
Deleting a conversation deletes its messages and everything remembered from it. For preferences you want applied reliably, use guidance. Memory is a convenience; guidance is an instruction.

​
Organize conversations

From a conversation’s menu you can rename it, pin it to the top of the sidebar, move it to a project, change its visibility, archive it or delete it. Titles are written for you from your first message, and you can change them. Conversations can be searched from the sidebar.

​
Who can see a conversation

​
Helpers

For research-heavy tasks the agent can start helpers: up to three at a time, each with read-only tools, working in parallel and reporting back. Helpers cannot edit documents, ask you questions or start helpers of their own. You can open a helper’s transcript from the conversation to see what it read.

​
Good habits

  • One subject per conversation. Long, wandering conversations cost more and drift. Start a new one when the subject changes.
  • Name the documents. Reference them rather than describing them.
  • Say what finished looks like: language, length, audience, format.
  • Stop and redirect. If the agent is heading the wrong way, stop it and say so. You do not need to wait for the turn to finish.